DERMARK introduces an advanced watermarking system that embeds traceable information in LLM-generated text while maintaining natural language quality.

• Creates multi-bit watermarks that identify specific LLM sources, not just whether AI generated the text
• Offers dynamic watermarking that adapts to different text characteristics
• Provides superior robustness against watermark erasure attacks
• Maintains high efficiency with minimal impact on generation quality

This research addresses critical security concerns around LLM misuse and copyright protection by enabling reliable attribution of AI-generated content, even when adversaries attempt to remove watermarks.

Original Paper: DERMARK: A Dynamic, Efficient and Robust Multi-bit Watermark for Large Language Models